Canadian AI infrastructure: What's available in 2026

The thing that actually stalled our procurement wasn't model quality. Every vendor let us test that for free. What stalled things was one line on a sub-processor table that nobody on the product side could explain for three business days.

We're a mid-size firm. Regulated enough that "we'll sort the privacy stuff out later" was never an option. We'd been asked to find a chat and document-review tool for about forty people who were, by that point, pasting client material into consumer AI tools on their own personal accounts. That's the real starting condition for most of these projects, I think — not a strategic initiative, just damage control that got formalized into a procurement.

The term that kept showing up in our own notes was "Canadian AI," mostly because our general counsel kept asking if there was a Canadian AI platform that did what the big US tools did, minus the jurisdiction question. There are more of these now than two years ago. Whether "more" means "enough" is a separate question. I don't think I fully answered it.

What we actually asked vendors

The list came together after the first demo went sideways. A vendor rep couldn't tell us, on the spot, where inference happened versus where data was stored — treated those as the same question. They're not. So every call after that, we asked the same things in the same order:

• Where is customer data stored, specifically, and is that in the contract or just in a sales deck?
• Where does inference actually run, including for any model tiers that differ from the default?
• Is there a US parent company or US investor with any claim on the data?
• What's in the sub-processor table, and does it include anything in the United States?
• What happens if the vendor's primary region goes down — does it fail over somewhere else, and where?
• Does our data train their models, and is that opt-out or architecturally not possible?

That last one mattered more than I expected. Two vendors could not give us a clean no.

The CLOUD Act question, and why it wasn't academic

Our counsel's position was narrow, which I appreciated after weeks of vague reassurance from sales teams. The concern wasn't abstract American overreach. It was that if a vendor has a US corporate parent, US courts can compel production of data the vendor holds, regardless of where the servers physically sit. That's the mechanism under the statute everyone calls the CLOUD Act.

Counsel's read — and I'm paraphrasing a conversation, not quoting a document — was that "the servers are in Canada" is an incomplete answer on its own. If the company controlling those servers is American, the legal exposure can still reach the data. That distinction is the one thing on our list that genuinely changed a decision. We moved two otherwise strong candidates out of contention once we confirmed they had US parents, even though their infrastructure pages said "hosted in Canada" in bold letters.

A sceptical colleague pushed back on this, reasonably: doesn't a mutual legal assistance treaty cover the Canada-US case anyway, so why does ownership matter so much? My understanding, again secondhand from counsel and not something I'd treat as settled law, is that MLAT requests run through courts in both countries and take time, while a CLOUD Act order served on a US parent doesn't require Canadian court involvement at all. Different process, different speed, different list of who gets notified. That's the gap we were trying to close, not a hypothetical one.

Augure's pitch was narrower than that, and I think more defensible for it: no US parent, no US investors, so customer conversations and documents aren't handled by a US-jurisdiction provider. That's a claim about customer content and corporate structure. It is not a claim that nothing in the stack ever touches US infrastructure, and when we pushed, the answer we got back matched that — email delivery and payment processing still run through standard US-based services, disclosed in the privacy policy rather than buried. I respected that more than the flat "fully Canadian, nothing touches the US" one competitor's sales engineer told us, which turned out to be false the moment we looked at their own sub-processor list.

Where Law 25 actually bit

For a Quebec-facing part of the business, the live question was section 17 of Law 25. It governs transfers of personal information outside Quebec and requires a privacy impact assessment before those transfers happen.

That assessment is where "where does inference run" stopped being theoretical. Augure runs some model tiers on Canadian infrastructure and others through EU partners, including during failover. They were upfront that it isn't all Canada, all the time. That turned out to be more useful for a section 17 assessment than a vague residency claim would have been, because we had to document which flows existed, not just get a yes or no on sovereignty.

Concretely, the assessment meant three things, in order: listing every category of personal information that would pass through the tool, mapping each category against where it would be processed, and then writing up the risk of the destination jurisdiction for each flow. The EU leg was the easy part to document, oddly — GDPR adequacy gave us language we could reuse almost directly. The harder paragraph was the US card-network and email processing, because that's not "AI processing" in any sense counsel cared about, but it's still a transfer outside Quebec and still had to appear in the assessment. Skipping it because it felt unrelated to the AI question would have been the actual compliance gap, not the EU inference.

PIPEDA mattered less than I expected going in. Our existing PIPEDA practices already assumed cross-border risk, so the AI tooling question didn't move that baseline much. I'll flag that as the thing that turned out not to matter as much as we braced for. We'd budgeted a lot of review time for PIPEDA gap analysis and ended up spending most of it on the Law 25 transfer question instead.

The model names turned out to be a rabbit hole

Early on we got hung up comparing model names and context windows across vendors, like that was the deciding axis. One flagship model advertised a 256K context window, with a premium tier going up to 1M. Those numbers sound decisive until you remember the actual use case is NDA triage and internal policy Q&A, where almost none of your documents get anywhere near that length.

We spent, roughly, a week on context-window comparisons that didn't move the final decision at all. I'd skip that step if I did this again.

What mattered more was plainer: persistent memory across sessions on the paid tier, so the tool didn't start from zero every new chat. On pricing, Augure's paid individual tier ran about C$20 a month with a weekly allowance in the range of 1,100 typical messages, and a C$80 tier above that added deep research agents and higher upload limits. That's a real number you can check against a quote, not a marketing abstraction, and it made the budget conversation shorter than I expected.

What I'd do differently

Bring the security reviewer in at the start, not after the shortlist. We were not sure, going in, how much weight to give corporate ownership relative to the technical security questions, and in hindsight it should have been first on the list, not third. Ownership determines which legal regime applies before you even get to encryption standards or SOC reports.

I also think we undervalued, early on, how much time a vendor built around Law 25 and PIPEDA saves on explanation. A Canadian AI platform doesn't satisfy those regulations on our behalf — it doesn't, and no vendor should claim that. But the questions get answered faster when the regulatory context was part of the design brief, not retrofitted after a US product got a Canadian landing page.

We didn't test every sovereign Canadian AI option on the market. I'd be lying if I called this review exhaustive. It was thorough enough for forty people and a general counsel who wanted a specific answer to a specific jurisdiction question. Next round, whenever that comes, I'd want to look at Augure again alongside whatever else has shown up by then.